optimized Rgba32.PixelOperations.PackFromVector4()

src/ImageSharp/Common/Extensions/SimdUtils.cs

@@ -2,13 +2,13 @@
 // Licensed under the Apache License, Version 2.0.
 
 using System;
+using System.Diagnostics;
 using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 
 namespace SixLabors.ImageSharp
 {
-    using System.Diagnostics;
-
     /// <summary>
     /// Various extension and utility methods for <see cref="Vector4"/> and <see cref="Vector{T}"/> utilizing SIMD capabilities
     /// </summary>
@@ -19,7 +19,6 @@ namespace SixLabors.ImageSharp
         /// </summary>
         public static readonly bool IsAvx2 = Vector<float>.Count == 8 && Vector<int>.Count == 8;
 
-        [Conditional("DEBUG")]
         internal static void GuardAvx2(string operation)
         {
             if (!IsAvx2)
@@ -79,11 +78,16 @@ namespace SixLabors.ImageSharp
 
             ref Vector<float> srcBase = ref Unsafe.As<float, Vector<float>>(ref source.DangerousGetPinnableReference());
             ref Octet.OfByte destBase = ref Unsafe.As<byte, Octet.OfByte>(ref dest.DangerousGetPinnableReference());
+            int n = source.Length / 8;
 
             Vector<float> magick = new Vector<float>(32768.0f);
             Vector<float> scale = new Vector<float>(255f) / new Vector<float>(256f);
 
-            int n = source.Length;
+            // need to copy to a temporal struct, because
+            // SimdUtils.Octet.OfUInt32 temp = Unsafe.As<Vector<float>, SimdUtils.Octet.OfUInt32>(ref x)
+            // does not work. TODO: This might be a CoreClr bug, need to ask/report
+            var temp = default(Octet.OfUInt32);
+            ref Vector<float> tempRef = ref Unsafe.As<Octet.OfUInt32, Vector<float>>(ref temp);
 
             for (int i = 0; i < n; i++)
             {
@@ -92,13 +96,10 @@ namespace SixLabors.ImageSharp
                 // return (uint8_t)u.i;
                 Vector<float> x = Unsafe.Add(ref srcBase, i);
                 x = (x * scale) + magick;
-
-                Vector<uint> u = Vector.AsVectorUInt32(x);
-
-                Octet.OfUInt32 ii = Unsafe.As<Vector<uint>, Octet.OfUInt32>(ref u);
+                tempRef = x;
 
                 ref Octet.OfByte d = ref Unsafe.Add(ref destBase, i);
-                d.LoadFrom(ref ii);
+                d.LoadFrom(ref temp);
             }
         }
 
@@ -118,11 +119,16 @@ namespace SixLabors.ImageSharp
 
             ref Vector<float> srcBase = ref Unsafe.As<float, Vector<float>>(ref source.DangerousGetPinnableReference());
             ref Octet.OfByte destBase = ref Unsafe.As<byte, Octet.OfByte>(ref dest.DangerousGetPinnableReference());
+            int n = source.Length / 8;
 
             Vector<float> magick = new Vector<float>(32768.0f);
             Vector<float> scale = new Vector<float>(255f) / new Vector<float>(256f);
 
-            int n = source.Length;
+            // need to copy to a temporal struct, because
+            // SimdUtils.Octet.OfUInt32 temp = Unsafe.As<Vector<float>, SimdUtils.Octet.OfUInt32>(ref x)
+            // does not work. TODO: This might be a CoreClr bug, need to ask/report
+            var temp = default(Octet.OfUInt32);
+            ref Vector<float> tempRef = ref Unsafe.As<Octet.OfUInt32, Vector<float>>(ref temp);
 
             for (int i = 0; i < n; i++)
             {
@@ -134,27 +140,80 @@ namespace SixLabors.ImageSharp
                 x = Vector.Min(x, Vector<float>.One);
 
                 x = (x * scale) + magick;
-
-                Vector<uint> u = Vector.AsVectorUInt32(x);
-
-                Octet.OfUInt32 ii = Unsafe.As<Vector<uint>, Octet.OfUInt32>(ref u);
+                tempRef = x;
 
                 ref Octet.OfByte d = ref Unsafe.Add(ref destBase, i);
-                d.LoadFrom(ref ii);
+                d.LoadFrom(ref temp);
             }
         }
 
-#pragma warning disable SA1132 // Do not combine fields
-        private static class Octet
+        // TODO: Replace these with T4-d library level tuples!
+        internal static class Octet
         {
+            [StructLayout(LayoutKind.Explicit, Size = 8 * sizeof(uint))]
             public struct OfUInt32
             {
-                public uint V0, V1, V2, V3, V4, V5, V6, V7;
+                [FieldOffset(0 * sizeof(uint))]
+                public uint V0;
+
+                [FieldOffset(1 * sizeof(uint))]
+                public uint V1;
+
+                [FieldOffset(2 * sizeof(uint))]
+                public uint V2;
+
+                [FieldOffset(3 * sizeof(uint))]
+                public uint V3;
+
+                [FieldOffset(4 * sizeof(uint))]
+                public uint V4;
+
+                [FieldOffset(5 * sizeof(uint))]
+                public uint V5;
+
+                [FieldOffset(6 * sizeof(uint))]
+                public uint V6;
+
+                [FieldOffset(7 * sizeof(uint))]
+                public uint V7;
+
+                public override string ToString()
+                {
+                    return $"[{this.V0},{this.V1},{this.V2},{this.V3},{this.V4},{this.V5},{this.V6},{this.V7}]";
+                }
             }
 
+            [StructLayout(LayoutKind.Explicit, Size = 8)]
             public struct OfByte
             {
-                public byte V0, V1, V2, V3, V4, V5, V6, V7;
+                [FieldOffset(0)]
+                public byte V0;
+
+                [FieldOffset(1)]
+                public byte V1;
+
+                [FieldOffset(2)]
+                public byte V2;
+
+                [FieldOffset(3)]
+                public byte V3;
+
+                [FieldOffset(4)]
+                public byte V4;
+
+                [FieldOffset(5)]
+                public byte V5;
+
+                [FieldOffset(6)]
+                public byte V6;
+
+                [FieldOffset(7)]
+                public byte V7;
+
+                public override string ToString()
+                {
+                    return $"[{this.V0},{this.V1},{this.V2},{this.V3},{this.V4},{this.V5},{this.V6},{this.V7}]";
+                }
 
                 public void LoadFrom(ref OfUInt32 i)
                 {

src/ImageSharp/PixelFormats/Rgba32.PixelOperations.cs

@@ -101,7 +101,6 @@ namespace SixLabors.ImageSharp
                 }
 
                 int remainder = count % Vector<uint>.Count;
-
                 int alignedCount = count - remainder;
 
                 if (alignedCount > 0)
@@ -117,6 +116,35 @@ namespace SixLabors.ImageSharp
                 }
             }
 
+            internal override void PackFromVector4(Span<Vector4> sourceVectors, Span<Rgba32> destColors, int count)
+            {
+                GuardSpans(sourceVectors, nameof(sourceVectors), destColors, nameof(destColors), count);
+
+                if (!SimdUtils.IsAvx2)
+                {
+                    base.PackFromVector4(sourceVectors, destColors, count);
+                    return;
+                }
+
+                int remainder = count % 2;
+                int alignedCount = count - remainder;
+
+                if (alignedCount > 0)
+                {
+                    Span<float> flatSrc = sourceVectors.Slice(0, alignedCount).NonPortableCast<Vector4, float>();
+                    Span<byte> flatDest = destColors.NonPortableCast<Rgba32, byte>();
+
+                    SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows(flatSrc, flatDest);
+                }
+
+                if (remainder > 0)
+                {
+                    // actually: remainder == 1
+                    int lastIdx = count - 1;
+                    destColors[lastIdx].PackFromVector4(sourceVectors[lastIdx]);
+                }
+            }
+
             /// <inheritdoc />
             internal override void PackFromRgba32(Span<Rgba32> source, Span<Rgba32> destPixels, int count)
             {

tests/ImageSharp.Benchmarks/Color/Bulk/PackFromVector4.cs

@@ -0,0 +1,65 @@
+// ReSharper disable InconsistentNaming
+namespace SixLabors.ImageSharp.Benchmarks.Color.Bulk
+{
+    using System.Numerics;
+
+    using BenchmarkDotNet.Attributes;
+
+    using SixLabors.ImageSharp.Memory;
+    using SixLabors.ImageSharp.PixelFormats;
+
+    [Config(typeof(Config.Short))]
+    public abstract class PackFromVector4<TPixel>
+        where TPixel : struct, IPixel<TPixel>
+    {
+        private Buffer<Vector4> source;
+
+        private Buffer<TPixel> destination;
+
+        [Params(16, 128, 512)]
+        public int Count { get; set; }
+
+        [GlobalSetup]
+        public void Setup()
+        {
+            this.destination = new Buffer<TPixel>(this.Count);
+            this.source = new Buffer<Vector4>(this.Count);
+        }
+
+        [GlobalCleanup]
+        public void Cleanup()
+        {
+            this.destination.Dispose();
+            this.source.Dispose();
+        }
+
+        [Benchmark(Baseline = true)]
+        public void PerElement()
+        {
+            Vector4[] s = this.source.Array;
+            TPixel[] d = this.destination.Array;
+            
+            for (int i = 0; i < this.Count; i++)
+            {
+                d[i].PackFromVector4(s[i]);
+            }
+        }
+
+        [Benchmark]
+        public void CommonBulk()
+        {
+            new PixelOperations<TPixel>().PackFromVector4(this.source, this.destination, this.Count);
+        }
+
+        [Benchmark]
+        public void OptimizedBulk()
+        {
+            PixelOperations<TPixel>.Instance.PackFromVector4(this.source, this.destination, this.Count);
+        }
+    }
+
+    public class PackFromVector4_Rgba32 : PackFromVector4<Rgba32>
+    {
+
+    }
+}

tests/ImageSharp.Tests/Common/SimdUtilsTests.cs

@@ -182,6 +182,21 @@ namespace SixLabors.ImageSharp.Tests.Common
             Assert.Equal(expected, actual);
         }
 
+        [Fact]
+        private void BulkConvertNormalizedFloatToByte_Step()
+        {
+            float[] source = {0, 7, 42, 255, 0.5f, 1.1f, 2.6f, 16f};
+            byte[] expected = source.Select(f => (byte)Math.Round(f)).ToArray();
+
+            source = source.Select(f => f / 255f).ToArray();
+
+            byte[] dest = new byte[8];
+
+            this.MagicConvert(source, dest);
+
+            Assert.Equal(expected, dest);
+        }
+
         private static byte MagicConvert(float x)
         {
             float f = 32768.0f + x;
@@ -189,6 +204,30 @@ namespace SixLabors.ImageSharp.Tests.Common
             return (byte)i;
         }
 
+        private void MagicConvert(Span<float> source, Span<byte> dest)
+        {
+            Vector<float> magick = new Vector<float>(32768.0f);
+            Vector<float> scale = new Vector<float>(255f) / new Vector<float>(256f);
+
+            Vector<float> x = source.NonPortableCast<float, Vector<float>>()[0];
+            
+            x = (x * scale) + magick;
+
+            SimdUtils.Octet.OfUInt32 ii = default(SimdUtils.Octet.OfUInt32);
+
+            ref Vector<float> iiRef = ref Unsafe.As<SimdUtils.Octet.OfUInt32, Vector<float>>(ref ii);
+
+            iiRef = x;
+
+            //SimdUtils.Octet.OfUInt32 ii = Unsafe.As<Vector<float>, SimdUtils.Octet.OfUInt32>(ref x);
+            
+            ref SimdUtils.Octet.OfByte d = ref dest.NonPortableCast<byte, SimdUtils.Octet.OfByte>()[0];
+            d.LoadFrom(ref ii);
+
+            this.Output.WriteLine(ii.ToString());
+            this.Output.WriteLine(d.ToString());
+        }
+
         private static void AssertEvenRoundIsCorrect(Vector<float> r, Vector<float> v)
         {
             for (int i = 0; i < Vector<float>.Count; i++)

tests/ImageSharp.Tests/PixelFormats/PixelOperationsTests.cs

@@ -12,10 +12,9 @@ namespace SixLabors.ImageSharp.Tests.PixelFormats
 {
     public partial class PixelOperationsTests
     {
-
-        public class Color32 : PixelOperationsTests<Rgba32>
+        public class Rgba32 : PixelOperationsTests<ImageSharp.Rgba32>
         {
-            public Color32(ITestOutputHelper output)
+            public Rgba32(ITestOutputHelper output)
                 : base(output)
             {
             }
@@ -26,19 +25,19 @@ namespace SixLabors.ImageSharp.Tests.PixelFormats
             [Fact]
             public void IsSpecialImplementation()
             {
-                Assert.IsType<Rgba32.PixelOperations>(PixelOperations<Rgba32>.Instance);
+                Assert.IsType<ImageSharp.Rgba32.PixelOperations>(PixelOperations<ImageSharp.Rgba32>.Instance);
             }
 
             [Fact]
             public void ToVector4SimdAligned()
             {
-                Rgba32[] source = CreatePixelTestData(64);
+                ImageSharp.Rgba32[] source = CreatePixelTestData(64);
                 Vector4[] expected = CreateExpectedVector4Data(source);
 
                 TestOperation(
                     source,
                     expected,
-                    (s, d) => Rgba32.PixelOperations.ToVector4SimdAligned(s, d, 64)
+                    (s, d) => ImageSharp.Rgba32.PixelOperations.ToVector4SimdAligned(s, d, 64)
                 );
             }
 
@@ -51,23 +50,23 @@ namespace SixLabors.ImageSharp.Tests.PixelFormats
                 int times = 200000;
                 int count = 1024;
 
-                using (Buffer<Rgba32> source = new Buffer<Rgba32>(count))
+                using (Buffer<ImageSharp.Rgba32> source = new Buffer<ImageSharp.Rgba32>(count))
                 using (Buffer<Vector4> dest = new Buffer<Vector4>(count))
                 {
                     this.Measure(
                         times,
                         () =>
                             {
-                                PixelOperations<Rgba32>.Instance.ToVector4(source, dest, count);
+                                PixelOperations<ImageSharp.Rgba32>.Instance.ToVector4(source, dest, count);
                             });
                 }
             }
         }
 
-        public class Argb : PixelOperationsTests<Argb32>
+        public class Argb32 : PixelOperationsTests<ImageSharp.PixelFormats.Argb32>
         {
             // For 4.6 test runner MemberData does not work without redeclaring the public field in the derived test class:
-            public Argb(ITestOutputHelper output)
+            public Argb32(ITestOutputHelper output)
                 : base(output)
             {
             }